# AI capex — X 热门讨论 (2026-09-26 06:57 UTC)
## @NinNinnin0306 (NinNin) · 09-26 04:08 · ♥72 ↻0 💬81 Physical AI doesn’t just need smarter models It needs better ground truth
A robot can have a powerful model, but it still needs to understand the environment around it
Where buildings are How spaces are structured What objects exist And how the physical world changes
This is the problem Vangrid is working on
Vangrid describes itself as a human collected spatial data layer designed to provide real time ground truth for world models and autonomous systems
Its infrastructure brings together:
- Multi-view ingestion - Edge-computed privacy - Cryptographic provenance - Enterprise Spatial API
The interesting part is the idea of a distributed, zero-capex sensor network, where real-world captures can contribute to a continuously growing spatial data layer
Vangrid is also experimenting with 3D capture bounties, where contributors can submit real world captures and accepted work is reconstructed into 3D models
If Physical AI is going to operate in the real world, it first needs a better way to understand it
That’s where spatial data becomes interesting
@vangrid_io > 引用 @NinNinnin0306: A map tells you where things are
But Physical AI needs more than a static map
It needs to understand how a place actually looks, what exists inside it, how objects relate to each other, and how the environment changes
That’s why spatial data is becoming increasingly important for autonomous systems
Vangrid is building a distributed spatial data layer around real world captures, turning everyday devices into potential sources of ground truth
The future of robotics may depend not only on smarter models, but on better representations of the world they operate in
@vangrid_io https://x.com/NinNinnin0306/status/2103698270128365601
## @tommykato (Tommy Kato 💫) · 09-26 04:00 · ♥65 ↻8 💬8 Everyone wants to know which of the Magnificent Seven AI kills. I think that’s the wrong question.
None of them need to disappear.
IBM didn’t disappear. Intel didn’t disappear. They just stopped being the scarce asset in the next computing cycle.
AI has three layers that matter:
Who owns the intelligence: models and data.
Who owns the factory: chips, power and cloud.
Who owns the last mile: the device, workplace, feed or agent where demand starts.
The obvious conclusion is that Apple and Microsoft have more to defend than most people think. Apple owns the device but rents much of the intelligence. Microsoft owns enormous enterprise distribution, but agents are coming directly for the software seat.
But there’s another side to this that I think matters even more.
The AI infrastructure trade isn’t seven different bets if they all depend on the same customer paying the bill.
NVIDIA, Broadcom, TSMC, the clouds and the AI labs can look diversified on a screen while ultimately depending on the same AI capex cycle.
Picks and shovels are only safe when the miners can pay.
That was one of the lessons of 2000. Equipment demand looked enormous until the financing underneath the customers broke.
I don’t think AI is 2000. The demand is real and the technology is real. But the financing still matters.
So I’ve stopped asking only which company wins AI.
I’m asking which companies still compound if the AI capex cycle slows for two years.
That gives me a very different list. https://x.com/tommykato/status/2103696299757355154
## @weakinstrument (Ryan Cummings) · 09-25 19:49 · ♥31 ↻8 💬3 Two additional things to stress:
1) Even if you make the most aggressive possible assumptions about incremental AI revenue, use the lowest discount rate possible, and assume all capex after 26 is canceled, to just *break even* the firms need to increase revenue 3-4x *next year* https://t.co/fjgSRzzIk9 > 引用 @weakinstrument: Me+@econJaredB are out w/ a new piece examining how much revenue is needed to justify hyperscaler AI capex. We estimate that incremental AI revenue is currently b/n $86-$188.1B for the hyperscalers, but they need $13.1-$18.7 *trillion* over the next decade for that to pencil out. https://t.co/h1INfoA7j5 https://x.com/weakinstrument/status/2103572560617353372